A new arXiv paper dissects the finite-step geometry of marginal-preserving velocity fields in generative flows, revealing why flow-matching models can sample in fewer steps — with direct implications for faster AI video and image synthesis.
A new deepfake detection approach breaks the monolithic 'fake' category into multiple subclasses, improving the model's ability to generalise to synthetic media produced by generators it never saw during training.
Zoom is rolling out deepfake detection for video meetings, bringing real-time synthetic media safeguards to one of the world's most-used enterprise conferencing platforms as impersonation attacks surge.
A new arXiv paper derives information-theoretic scaling laws for generative diffusion training, linking data, compute, and model capacity to the mutual information a diffusion model can capture — with direct implications for AI image and video generation.
Opera has proposed a new web standard that lets websites deploy and run AI models locally in users' browsers, pushing inference on-device for privacy, speed, and offline capability.
A technical walkthrough of the LLM inference pipeline — from tokenization and embeddings to attention layers and decoding — explaining exactly what unfolds between your prompt and the model's response.
Pennsylvania has enacted a law requiring clear disclosures on AI-generated advertising, joining a growing patchwork of US state rules targeting synthetic media and deceptive deepfake content in commercial and political messaging.
Zoom unveiled a suite of AI updates including a Unified Context Layer, My Notes workflows, and built-in deepfake detection—bringing synthetic media defenses directly into enterprise video conferencing.
Speculative decoding uses a small draft model to predict tokens that a larger model verifies in parallel, delivering up to 3x faster inference without sacrificing output quality. Here's how the technique works and why it matters for generative AI.
Jack Dorsey's Block has unveiled an AI voice model from its Dessa lab designed to sound like an awkward but endearing new friend—pushing conversational speech synthesis toward imperfect, human-like naturalness rather than polished perfection.
A technical breakdown of how modern reasoning models deliberate, backtrack, and self-correct mid-generation — and why tokens like 'wait' signal measurable shifts in an AI's internal reasoning trajectory.
A technical deep dive into the mechanics behind AI image generation — covering diffusion models, latent space, and the newer flow matching approach — and why quirks like clocks always showing 10:10 reveal how these systems actually learn.